README.md

June 26, 2025 · View on GitHub

[ICCV 2025] Timestep-Aware Diffusion Model for Extreme Image Rescaling

     

Ce Wang1, Zhenyu Hu1, Zhenzhong Chen1, Wanjie Sun1 1School of Remote Sensing and Information Engineering, Wuhan University

📚 Table of Contents

👁️ Visual Results

Results on DIV2K and CLIC2020

Results on Urban100

Results on DIV8K

⚙️ Installation

# Clone this repository
git clone https://github.com/wwangcece/TADM.git

# Create a conda environment with Python >= 3.9
conda create -n TADM python=3.9
conda activate TADM

# Install required packages
pip install -r requirements.txt

🧬 Pretrained Models

Download the pretrained models from the link below and place them in the checkpoints/ directory:

Download from Google Drive

📊 Dataset

  1. Refer to the BasicSR dataset preparation guide to prepare high-resolution datasets.
  2. Use the script src/get_z.py to encode high-resolution images into latent features and save them as .npy files.

⚔️ Inference

  1. Modify the validation dataset configuration in configs/tadm_test.yaml and update the pretrained model path in run_inference.sh.
  2. Run the inference script:
sh run_inference.sh

🌟 Training

  1. Modify the training dataset configuration in configs/tadm_train.yaml and update settings in run_training_dfrm.sh.
  2. Train the feature rescaling module:
sh run_training_dfrm.sh
  1. Then modify run_training_tadm.sh as needed and train the TADM model:
sh run_training_tadm.sh

📖 Citation

If you find this work helpful, please consider citing:

@misc{wang2025timestepawarediffusionmodelextreme,
  title={Timestep-Aware Diffusion Model for Extreme Image Rescaling},
  author={Ce Wang and Zhenyu Hu and Wanjie Sun and Zhenzhong Chen},
  year={2025},
  eprint={2408.09151},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2408.09151}
}

🙏 Acknowledgements

This project is based on S3Diff. We thank the authors for their excellent work.

📨 Contact

If you have any questions, feel free to reach out to: Ce Wang — cewang@whu.edu.cn